Files
leonarski_f 4809c1622c
Build Packages / build:portable:linux-x86_64:nocuda (push) Successful in 16m43s
Build Packages / build:jfjoch:ubuntu2404:nocuda (push) Successful in 28m46s
Build Packages / build:jfjoch:rocky8:cuda-sls9 (push) Successful in 32m5s
Build Packages / build:jfjoch:rocky9:cuda (push) Successful in 39m44s
Build Packages / build:portable:macos-arm64:nocuda (push) Successful in 5m7s
Build Packages / Create release (push) Successful in 8s
Build Packages / Build documentation (push) Successful in 49s
Build Packages / build:portable:linux-x86_64:cuda (push) Successful in 22m30s
Build Packages / build:jfjoch:rocky8:nocuda (push) Successful in 31m54s
Build Packages / build:rugnux:linux-aarch64:cuda (push) Successful in 11m29s
Build Packages / build:portable:windows-x86_64:nocuda (push) Successful in 23m51s
Build Packages / build:jfjoch:ubuntu2204:nocuda (push) Successful in 26m7s
Build Packages / build:jfjoch:ubuntu2204:cuda (push) Successful in 29m38s
Build Packages / build:portable:windows-x86_64:cuda (push) Successful in 28m6s
Build Packages / HDF5 consumer tests (DIALS, XDS) (push) Successful in 29m44s
Build Packages / build:jfjoch:rocky9:nocuda (push) Successful in 33m22s
Build Packages / Generate python client (push) Successful in 47s
Build Packages / build:jfjoch:rocky9:cuda-sls9 (push) Successful in 34m6s
Build Packages / Unit tests (push) Successful in 1h2m47s
Build Packages / build:jfjoch:rocky8:cuda (push) Successful in 35m59s
Build Packages / build:jfjoch:ubuntu2404:cuda (push) Successful in 34m8s
Build Packages / Validation Battery (push) Successful in 38m31s
v1.0.0-rc.187 (#98)
* Rugnux: Refinement several times faster, one protocol on the CPU and the GPU, parts of it on the Apple GPU; its random draws are the same on every platform.
* Rugnux: Refinement follows common practice: the free set decides only the restraint weights and the displacement model (isotropic, TLS or anisotropic, by Hamilton's test), everything else is decided on the working set; lower R-free for placed and predicted models.
* Rugnux: Restraint dictionaries via `--restraints` (or among the inputs) and more built-in ligands; a model whose space group is a subgroup of the data's is refined in its own group; reflections plus a model select `--mode refine` on their own.
* Rugnux: `-C` fixes the crystal system on the axes as typed; `-S` alone fixes the space group only, the cell is still found and reduced; a space group or cell stored in the input file is ignored, with a line in the log.
* Rugnux: More reliable multi-sweep merging; `--mode scale` merges several stills runs as one; `_plot.txt` and the report name each input; a second crystal form is reported for stills given a cell.
* jfjoch_broker (experimental): Live Rugnux on FPGA detectors (JUNGFRAU) takes the same path as a DECTRIS stream, rotation and serial alike, model step included, with analysis threads per GPU; frames waiting for the analysis can be viewed; no host stage drops a frame - a source faster than the analysis is held at the card.
* jfjoch_broker: Summation of more than 256 frames per image is split between the card and the host (`summation_mode` is then `hybrid`), and live Rugnux runs on such collections; pedestal runs take a receiver of their own.
* jfjoch_broker (experimental): `stills_downsample` analyses 1 of every N serial images (chosen by a hash; the first 2000 all; the others written without analysis; reset to 1 by `/initialize`); `max_spot_count` left out means 250 for per-image indexing and up to 1000 for live stills; the live status gives the merge before the model work, the fitted resolution and the anomalous peak height of each unexplained blob.
* jfjoch_broker: Without a `pcie` entry the broker runs on a mock of the card instead of the HLS simulation; its spot, ROI and azimuthal results are zero then.
* Web frontend: Redesigned Rugnux live page; Vite 8, TypeScript 6, Mol* 5.13 and dependency security updates.
* Tools: `jfjoch_fpga_stills_test` measures live stills on the cards with a stored dataset; `jfjoch_fpga_test` loses the HLS-simulation option.
* Other: FFTW no longer fetched or linked (the CPU FFT indexer keeps the name `FFTW`); PCIe driver now GPL-2.0-or-later and third-party notices complete; several crash fixes.

Reviewed-on: #98
Co-authored-by: Filip Leonarski <filip.leonarski@psi.ch>
2026-11-19 20:49:28 +01:00

91 lines
5.7 KiB
C++

// SPDX-FileCopyrightText: 2026 Filip Leonarski, Paul Scherrer Institute <filip.leonarski@psi.ch>
// SPDX-License-Identifier: GPL-3.0-only
#pragma once
// The molecular-replacement search on a GPU (CUDA builds only; the header itself needs no CUDA): the
// rotation function and the translation function of MolecularReplacement.h, the same functions as the
// CPU computes - the same trilinear interpolation of the same molecular transform, the same bins, the
// same coefficients on the same grid - in float on the device, with cuFFT for pocketfft. The two agree to
// rounding, not bit for bit. Every reduction runs in a fixed order and nothing is summed with an
// atomic, so the GPU is deterministic on its own.
#include <complex>
#include <functional>
#include <memory>
#include <vector>
#include "MolecularReplacement.h"
// CropHalfL(MapToFPhi(map), reach) (ModelFFT.h) with cuFFT on the calling thread's current device: the same
// transform, the same half-l layout and conventions, to float rounding. For the molecular transform of a large
// model, whose box is hundreds of millions of points. MapToFPhiGPUBytes: the device memory it takes while it runs.
size_t MapToFPhiGPUBytes(const gemmi::Grid<float> &map, int reach);
gemmi::FPhiGrid<float> MapToFPhiGPU(const gemmi::Grid<float> &map, int reach);
struct MolecularReplacementGPUState;
struct MolecularReplacementGPUGroup;
struct TranslationSetupGPU;
class MolecularReplacementGPU {
public:
// Uploads the molecular transform, on the calling thread's current device; `packed`: with each point's
// interpolation corners together, read from two places instead of four, for more memory (TransformBytes)
// and to the same bits. Throws JFJochException on a CUDA failure, as every call below does.
explicit MolecularReplacementGPU(const MolecularTransform &ft, bool packed = false);
~MolecularReplacementGPU();
std::vector<double> RotationFunction(const SearchReflections &refl, const gemmi::UnitCell &cell,
const gemmi::SpaceGroup &sg, const std::vector<gemmi::Mat33> &rotations);
// The translation function of one group at each of `rotations`; with `known`, beside a known partial
// structure (KnownStructureFactors). `prepared`: Prepare() of this setup, for a setup used again and again;
// made for this call otherwise.
std::vector<TranslationPeak> TranslationFunction(const TranslationSetup &setup,
const std::vector<gemmi::Mat33> &rotations,
const std::vector<std::complex<double>> *known = nullptr,
TranslationSetupGPU *prepared = nullptr);
// What the translation function needs of `setup`, worked out once on the host; with `known_part`, for the
// function beside a known partial structure.
std::unique_ptr<TranslationSetupGPU, void (*)(TranslationSetupGPU *)> Prepare(const TranslationSetup &setup,
bool known_part);
// The device memory the transform is held in, plain or packed; while the packed one is made, both.
static size_t TransformBytes(const MolecularTransform &ft, bool packed);
// The device memory a call takes for as long as it runs, and gives back when it returns: RotationFunction()
// of `rotations` against `refl`'s reflections in `sg`, and TranslationFunction() of `rotations` (with
// `known`, beside a known partial structure).
static size_t RotationBytes(const SearchReflections &refl, const gemmi::SpaceGroup &sg, size_t rotations);
static size_t TranslationBytes(const TranslationSetup &setup, size_t rotations, bool known);
private:
std::unique_ptr<MolecularReplacementGPUState> state_;
};
// The search on one card: the transform uploaded on `device`.
struct MolecularReplacementGPUOnCard {
int device = 0;
MolecularReplacementGPU *gpu = nullptr;
};
// What a function's slice holds on its card while it runs: called on the slice's thread, with the card and the
// bytes its buffers take, and kept until the slice is done. Empty: nothing is held.
using MolecularReplacementGPUHold = std::function<std::shared_ptr<void>(int device, size_t bytes)>;
// The rotation and translation functions with the orientations split over `cards`: the k-th card scores the
// k-th of cards.size() contiguous slices of `rotations`, on a thread of its own beside the others, and the
// results are put back in the order of `rotations`. Every orientation is scored by itself - one block of the
// rotation function, one map of the translation function - and nothing is summed across orientations, so the
// result is that of one card, bit for bit, on any number of cards.
std::vector<double> SplitRotationFunction(const std::vector<MolecularReplacementGPUOnCard> &cards,
const SearchReflections &refl, const gemmi::UnitCell &cell,
const gemmi::SpaceGroup &sg, const std::vector<gemmi::Mat33> &rotations,
const MolecularReplacementGPUHold &hold = {});
std::vector<TranslationPeak> SplitTranslationFunction(const std::vector<MolecularReplacementGPUOnCard> &cards,
const TranslationSetup &setup,
const std::vector<gemmi::Mat33> &rotations,
const std::vector<std::complex<double>> *known,
TranslationSetupGPU *prepared,
const MolecularReplacementGPUHold &hold = {});